We present a novel experiment and perspective on human adaptive control. In our task, participants repeatedly adjust, zero, one or two variables with the goal of controlling a third variable, targeting a moving reward region. Across tasks, we vary the function that maps controls to target variables, and use computational modeling to examine how participants represent and solve the tasks. We find that, while broadly successful, participants fall back on assuming locally linear monotonic relationships, while also taking control actions that are conservative, preferring to adjust one variable rather than both relative to their previous action. We suggest that this allows for robust performance even when interacting with nonlinear non-monotonic functions.